Re: [PATCH v2 4/5] KVM: selftests: Add a test for KVM_RUN+rseq to detect task migration bugs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




----- On Aug 26, 2021, at 7:54 PM, Sean Christopherson seanjc@xxxxxxxxxx wrote:

> On Thu, Aug 26, 2021, Mathieu Desnoyers wrote:
>> ----- On Aug 25, 2021, at 8:51 PM, Sean Christopherson seanjc@xxxxxxxxxx wrote:
>> >> >> +		r = sched_setaffinity(0, sizeof(allowed_mask), &allowed_mask);
>> >> >> +		TEST_ASSERT(!r, "sched_setaffinity failed, errno = %d (%s)",
>> >> >> +			    errno, strerror(errno));
>> >> >> +		atomic_inc(&seq_cnt);
>> >> >> +
>> >> >> +		CPU_CLR(cpu, &allowed_mask);
>> >> >> +
>> >> >> +		/*
>> >> >> +		 * Let the read-side get back into KVM_RUN to improve the odds
>> >> >> +		 * of task migration coinciding with KVM's run loop.
>> >> > 
>> >> > This comment should be about increasing the odds of letting the seqlock
>> >> > read-side complete. Otherwise, the delay between the two back-to-back
>> >> > atomic_inc is so small that the seqlock read-side may never have time to
>> >> > complete the reading the rseq cpu id and the sched_getcpu() call, and can
>> >> > retry forever.
>> > 
>> > Hmm, but that's not why there's a delay.  I'm not arguing that a livelock isn't
>> > possible (though that syscall would have to be screaming fast),
>> 
>> I don't think we have the same understanding of the livelock scenario. AFAIU the
>> livelock
>> would be caused by a too-small delay between the two consecutive atomic_inc()
>> from one
>> loop iteration to the next compared to the time it takes to perform a read-side
>> critical
>> section of the seqlock. Back-to-back atomic_inc can be performed very quickly,
>> so I
>> doubt that the sched_getcpu implementation have good odds to be fast enough to
>> complete
>> in that narrow window, leading to lots of read seqlock retry.
> 
> Ooooh, yeah, brain fart on my side.  I was thinking of the two atomic_inc() in
> the
> same loop iteration and completely ignoring the next iteration.  Yes, I 100%
> agree
> that a delay to ensure forward progress is needed.  An assertion in main() that
> the
> reader complete at least some reasonable number of KVM_RUNs is also probably a
> good
> idea, e.g. to rule out a false pass due to the reader never making forward
> progress.

Agreed.

> 
> FWIW, the do-while loop does make forward progress without a delay, but at ~50%
> throughput, give or take.

I did not expect absolutely no progress, but a significant slow down of
the read-side.

> 
>> > but the primary motivation is very much to allow the read-side enough time
>> > to get back into KVM proper.
>> 
>> I'm puzzled by your statement. AFAIU, let's say we don't have the delay, then we
>> have:
>> 
>> migration thread                             KVM_RUN/read-side thread
>> -----------------------------------------------------------------------------------
>>                                              - ioctl(KVM_RUN)
>> - atomic_inc_seq_cst(&seqcnt)
>> - sched_setaffinity
>> - atomic_inc_seq_cst(&seqcnt)
>>                                              - a = atomic_load(&seqcnt) & ~1
>>                                              - smp_rmb()
>>                                              - b = LOAD_ONCE(__rseq_abi->cpu_id);
>>                                              - sched_getcpu()
>>                                              - smp_rmb()
>>                                              - re-load seqcnt/compare (succeeds)
>>                                                - Can only succeed if entire read-side happens while the seqcnt
>>                                                  is in an even numbered state.
>>                                              - if (a != b) abort()
>>   /* no delay. Even counter state is very
>>      short. */
>> - atomic_inc_seq_cst(&seqcnt)
>>   /* Let's suppose the lack of delay causes the
>>      setaffinity to complete too early compared
>>      with KVM_RUN ioctl */
>> - sched_setaffinity
>> - atomic_inc_seq_cst(&seqcnt)
>> 
>>   /* no delay. Even counter state is very
>>      short. */
>> - atomic_inc_seq_cst(&seqcnt)
>>   /* Then a setaffinity from a following
>>      migration thread loop will run
>>      concurrently with KVM_RUN */
>>                                              - ioctl(KVM_RUN)
>> - sched_setaffinity
>> - atomic_inc_seq_cst(&seqcnt)
>> 
>> As pointed out here, if the first setaffinity runs too early compared with
>> KVM_RUN,
>> a following setaffinity will run concurrently with it. However, the fact that
>> the even counter state is very short may very well hurt progress of the read
>> seqlock.
> 
> *sigh*
> 
> Several hours later, I think I finally have my head wrapped around everything.
> 
> Due to the way the test is written and because of how KVM's run loop works,
> TIF_NOTIFY_RESUME or TIF_NEED_RESCHED effectively has to be set before KVM
> actually
> enters the guest, otherwise KVM will exit to userspace without touching the
> flag,
> i.e. it will be handled by the normal exit_to_user_mode_loop().
> 
> Where I got lost was trying to figure out why I couldn't make the bug reproduce
> by
> causing the guest to exit to KVM, but not userspace, in which case KVM should
> easily trigger the problematic flow as the window for sched_getcpu() to collide
> with KVM would be enormous.  The reason I didn't go down this route for the
> "official" test is that, unless there's something clever I'm overlooking, it
> requires arch specific guest code, and ialso I don't know that forcing an exit
> would even be necessary/sufficient on other architectures.
> 
> Anyways, I was trying to confirm that the bug was being hit without a delay,
> while
> still retaining the sequence retry in the test.  The test doesn't fail because
> the
> back-to-back atomic_inc() changes seqcnt too fast.  The read-side makes forward
> progress, but it never observes failure because the do-while loop only ever
> completes after another sched_setaffinity(), never after the one that collides
> with KVM because it takes too long to get out of ioctl(KVM_RUN) and back to the
> test.  I.e. the atomic_inc() in the next loop iteration (makes seq_cnt odd)
> always
> completes before the check, and so the check ends up spinning until another
> migration, which correctly updates rseq.  This was expected and didn't confuse
> me.
> 
> What confused me is that I was trying to confirm the bug was being hit from
> within
> the kernel by confirming KVM observed TIF_NOTIFY_RESUME, but I misunderstood
> when
> TIF_NOTIFY_RESUME would get set.  KVM can observe TIF_NOTIFY_RESUME directly,
> but
> it's rare, and I suspect happens iff sched_setaffinity() hits the small window
> where
> it collides with KVM_RUN before KVM enters the guest.
> 
> More commonly, the bug occurs when KVM sees TIF_NEED_RESCHED.  In that case, KVM
> calls xfer_to_guest_mode_work(), which does schedule() and _that_ sets
> TIF_NOTIFY_RESUME.  xfer_to_guest_mode_work() then mishandles TIF_NOTIFY_RESUME
> and the bug is hit, but my confirmation logic in KVM never fired.
> 
> So there are effectively three reasons we want a delay:
> 
>  1. To allow sched_setaffinity() to coincide with ioctl(KVM_RUN) before KVM can
>     enter the guest so that the guest doesn't need an arch-specific VM-Exit source.
> 
>  2. To let ioctl(KVM_RUN) make its way back to the test before the next round
>     of migration.
> 
>  3. To ensure the read-side can make forward progress, e.g. if sched_getcpu()
>     involves a syscall.
> 
> 
> After looking at KVM for arm64 and s390, #1 is a bit tenuous because x86 is the
> only arch that currently uses xfer_to_guest_mode_work(), i.e. the test could be
> tweaked to be overtly x86-specific.  But since a delay is needed for #2 and #3,
> I'd prefer to rely on it for #1 as well in the hopes that this test provides
> coverage for arm64 and/or s390 if they're ever converted to use the common
> xfer_to_guest_mode_work().

Now that we have this understanding of why we need the delay, it would be good to
write this down in a comment within the test.

Does it reproduce if we randomize the delay to have it picked randomly from 0us
to 100us (with 1us step) ? It would remove a lot of the needs for arch-specific
magic delay value.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com



[Index of Archives]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Share Photos]     [IDE]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux ATA RAID]     [Samba]     [Device Mapper]

  Powered by Linux